Evaluating the Relationship Between MR Image Quality Measures and Deep Learning-Based Brain Tumor Segmentation Accuracy
R Muthusivarajan1*, A Celaya2, J Yung3, S Viswanath4, D Marcus5, C Chung6, D Fuentes7, (1) UT MD Anderson Cancer Center, ,,(2) MD Anderson Cancer Center, Houston, TX, (3) UT MD Anderson Cancer Center, Houston, TX, (4) Case Western Reserve University, ,,(5) ,,,(6) The University of Texas MD Anderson Cancer Center, Houston, TX, (7) UT MD Anderson Cancer Center, Houston, TX
Presentations
PO-GePV-M-293 (Sunday, 7/10/2022) [Eastern Time (GMT-4)]
ePoster Forums
Purpose: To evaluate the relationship between MR image quality measures and automated brain tumor segmentation accuracy.
Methods: For this study 369 multimodal MRI scans native (T1), post contrast T1-weighted (T1Gd), T2-weighted (T2), and T2 Fluid attenuated inversion recovery (T2-FLAIR) from BraTS 2020 cohort were considered. All four MRI scans were evaluated through the open-source quality control tool MRQy, to yield 13 image quality measures (IQMs) per scan. Five-fold cross validation performance of a 3D DenseNet model was evaluated in terms of dice overlap for each of whole tumor (WT), tumor core (TC) and enhancing tumor (ET) regions. Pearson correlation coefficient was computed between WT dice and IQM measures of the four scans. MRI scans were grouped as “better” quality (BQ) and “worse” quality (WQ) based on IQM measure. For each IQM quality grouping, BQ MRI scans were used to train the DenseNet model with validation on WQ scans. The inverse experiment was also conducted (train on WQ, validated on BQ).
Results: Based on correlation studies, seven T2-FLAIR IQMs were selected for further quality analysis. Training on BQ scans yielded significantly improved tumor segmentation accuracy for 2 IQMs, coefficient of variance (CV) and coefficient of joint variation (CJV), compared to training on WQ scans and 5-fold cross-validation. This direct correlation trend was further validated on the independent test set from the BraTS 2021 cohort. By contrast, training on WQ scans as defined by signal to noise ratios (SNR4), contrast to noise ratio (CNR) and peak SNR (PSNR) IQMs yielded better performance than BQ scans.
Conclusion: Our results suggest a significant correlated relationship may exist between specific MR image quality measures (CV, CJV) and DenseNet-based tumor segmentation performance, using multi-parametric MRIs. Selecting MRI scans for model training based on MR IQMs may yield more accurate and generalizable models compared to alternative strategies.
Funding Support, Disclosures, and Conflict of Interest: This work is supported through the MD Anderson Strategic Research Initiative Development (STRIDE) Program Tumor Measurement Initiative
Keywords
Image Analysis, Image Processing, Quality Control
Taxonomy
Not Applicable / None Entered.
Contact Email